AI, healthcare and the meaning of a measurement

Measurement is a starting point

As a pulmonary function analyst, I work with physiology, measurements and interpretation. A measurement gains meaning through the quality of its acquisition, its relationship to other values and the patient’s context.

That combination makes AI interesting to me. On PFTinfo I write about AI in pulmonary function testing. This page brings together the central questions from three articles for readers approaching healthcare from an AI perspective.

From individual values to patterns

Pulmonary function tests produce more than a few numbers. Curves contain shape information; different measurement techniques describe a more complex picture together. My article about AI and pulmonary function explores support for interpretation and quality assessment.

The question is how a model can help a professional evaluate patterns. Its output needs to be considered alongside measurement quality, limitations and clinical context. A convincing prediction alone is not a clinical conclusion.

What might existing signals contain?

From breathing curves to insight explores the potential value of capnograms and historical pulmonary function data. What could we learn from signal shape and relationships between measurements?

Practical research questions follow: can raw signals be exported, are acquisition conditions known and are reliable labels available? Historically collected data is not automatically a suitable training set.

Reuse of health data also requires establishing the applicable privacy, research and processing conditions. This page does not provide a legal assessment.

The connection to local AI

My personal Diet Coach project approaches a related question from another direction: how can a system use personal context while keeping processing in an environment under my control?

The project distinguishes observations, estimates, corrections and confirmed knowledge. That distinction matters when AI uses data over time. An error should not silently become an established part of its memory.

Questions I want to keep asking

• What evidence supports a model’s output, and can it be traced?
• How is uncertainty expressed, and when is professional review needed?
• What value does the system add to an actual workflow?
• Which data is necessary, and where is it processed?

My interest lies at the intersection of these questions: models, the systems around them and the practice in which their outputs acquire meaning.

This is an English editorial adaptation of my Dutch PFTinfo articles. It is an exploration, not a clinical guideline or medical application. The original articles are linked below.